Research On Image Segmentation Algorithm Of Pulmonary Nodules Based On Deep Learning

In medical imaging,lung cancer is mostly manifested as nodular with irregular shape and blurred edges.It is the key means to prevent the deterioration of lesions and improve the cure rate to distinguish benign and malignant pulmonary nodules with the help of medical images and make a diagnosis and treatment plan in time.It can provide an important reference for the diagnosis of lung tumors to segment pulmonary nodules from lung CT images.In recent years,the research of Deep Learning in medical image processing is more and more exhaustive,it has stronger adaptability and higher efficiency than the traditional segmentation methods in organ lesion segmentation,but there is still a problem of low accuracy.Under this background,deep-learning algorithms for lung nodule segmentation in CT image are constructed.The main research contents are as follows:Due to the complex imaging features of pulmonary nodules,the existing neural network model is not ideal for segmentation.To solve this problem,based on the UNet,classical algorithm model in biomedical image segmentation,the research constructs an encoder-decoder network model(DB-UNet)composed of dense block.Dense connections are introduced into each convolution module of U-Net to improve the accuracy of segmentation from the perspective of enhancing feature transfer and reuse.At the same time,to reduce the problem of gradient disappearance,Batch Normalization layers are inserted behind each convolution layer to improve the data distribution in networks layers.Experimental data show that the Dice coefficient of test set segmentation by this model reaches 0.849.Ablation experiments show that,dense connection structures effectively improve the segmentation accuracy of pulmonary nodules.The diagram of curves of loss in Validation Set show that,BN enhances the effectiveness of data transmission in layers of network,and accelerate network convergence.In view of the memory overhead caused by feature splicing in the network,an optimization method of DB-UNet segmentation is proposed.The number of skip connections in each Dense Block are reduced to one,and a skip connection convolution block is constructed.In order to make the model maintain high segmentation performance while reducing feature reuse,the channel attention modules are introduced into the encoder and decoder of the model,so that the layers of network can focus on the feature channel which containing pulmonary nodule information.A U-Net of Skip Connection Blocks based on channel attention mechanism(CAM-SB-UNet)was designed and constructed.In the experimental results,the recall rate and Jaccard coefficient of the segmentation reach 0.85 and 0.746 respectively.It shows that the Channel Attention module improves the efficiency of feature utilization,and the segmentation performance of the network is improved compared with DB-UNet,and reduces the resource overhead to a certain extent.